Health economics evaluation of artificial intelligence in the field of oncology: a scoping review

Abstract

Background

The ageing and growing population, rising cancer rate, and increased healthcare demand drive higher medical costs. Integrating Artificial Intelligence (AI) into oncology revolutionizes cancer diagnosis, treatment, and management, promising enhanced efficiency and precision. Despite potential for healthcare cost savings from implementing AI system in healthcare, its successful integration depends on stakeholder acceptance, particularly physician adoption and integration into routine clinical practice.

Objectives

This scoping review aimed to systematically assess recent literature on health economic evaluations (HEEs) of artificial AI applications in oncology and explore the economic risks and benefits of integrating AI systems into oncology.

Methodology

A scoping review was conducted using PRISMA-ScR guidelines across databases, including PubMed, Scopus, and Google Scholar for articles published between January 1, 2019, and June 30, 2024, in English. Eligible studies focused on the economic aspects of AI applications in oncology, including cost-effectiveness analyses, budget impact studies, and evaluations of economic benefits in clinical practice. Two independent reviewers used CHEERS-AI and Philips checklists for data extraction, quality assessment and analysis.

Results

Out of 870 studies identified,12 studies were selected which focused on colorectal (5), breast (2), lung (2), cervical (1), and prostate (1) cancers. Most emphasis is on early-stage care (N = 10,83%) and cost-effectiveness analysis (N = 9,75%) was predominant along with Markov model being the most common approach. The studies used a healthcare system perspective (N = 6,50%) and examined AI's cost-effectiveness in medical imaging (N = 5,42%) and biomarkers (N = 4,33%). Time horizons varied from less than a year to a lifetime, with most applying a 3% discount rate. AI demonstrated economic benefits, improved diagnostic sensitivity,potential cost reduction, workflow efficiency, and treatment optimization while presenting risks like reimbursement challenges, data security concerns, and potential error costs.

Conclusion

The scoping review highlights the necessity for thorough health economic evaluations of AI integration in oncology. Although AI technologies are cost-effective, there is a gap in user perspectives and considerations of health equity, regulation, and ethics. To maximize its benefits, future research should include a comprehensive economic evaluation of a more diverse population for the effective adoption of AI into the healthcare system.

Supplementary Information

The online version contains supplementary material available at 10.1186/s13561-026-00751-x.

Background

Cancer was the second leading cause of death globally with estimating 20 million new cancer cases and 9.7 million deaths worldwide in 2022 [1]. Lung, prostate, colorectal, stomach and liver cancers are the most prevalent cancers among men while breast, colorectal, lung, cervical and thyroid cancers are commonest in women. The global cancer burden is increasing, placing significant physical, emotional, and financial strain on individuals, families, communities, and health systems [1]. The cost of medical care is rising due to several factors, including an ageing and expanding population, rising illness rates, an increase in the number of patients utilizing healthcare services, and rising expenses [2]. A quarter of all healthcare expenditures in the US were wasted [3]. Overtreatment, inadequate health care delivery, and other preventable and correctable system flaws are the main contributors to this expense [4]. Artificial intelligence (AI) has rapidly advanced in healthcare due to growing interest, enhanced processing power, improved algorithms, and extensive digitization of health data, revolutionizing patient-centered care and presenting both new opportunities and challenges for clinical practice [5–7].

The strategic deployment of AI systems in healthcare can mitigate rising healthcare costs by eliminating inefficiencies in non-clinical administration and enhancing the precision of clinical resource management [8]. However, the adoption of AI requires support from stakeholders, particularly the doctors who will be using the system [8]. Specifically in the field of oncology, AI has emerged as a transformative technology, offering promising advancements in the accuracy and efficiency of cancer screening, detection, and diagnosis, accelerating drug discovery, facilitating precision treatment, and improving management, care, and surveillance [9]. Recent developments in AI include the Mirai machine learning model for improved mammography interpretation in breast cancer, the Sybil model for analyzing low-dose chest CT scans in lung cancer, and automation tools like Curate AI for cancer care [10–12].

According to Global Market Insights, the AI in Oncology market, encompassing software solutions, hardware, services, and applications across various cancer types (breast, lung, brain, prostate, colorectal) and fields (diagnostics, immunotherapy, radiation therapy, R&D), was valued at approximately USD 1.9 billion in 2023. By 2032, this market is projected to reach a value of more than USD 17.9 billion [13]. Despite this rapid growth, there is a notable gap in the literature focus on the economic modelling of the true impact of AI in oncology. Most papers published around AI applications in oncology centers on accuracy and precision of prediction [14, 15]. Achieving high prediction accuracy and distinct correlations between features of the patient or image and outcomes does not, however, guarantee clinical efficacy or broad acceptance. Furthermore, clinical relevance is not always indicated by the area under the receiver operating characteristic curve, which is a common metric used to evaluate detection tasks [14, 15]. After the Covid 19 pandemic, with digital transformation demanded advancement in AI technology and its impact on healthcare economic structures. Earlier studies on foundational AI model might differ from existing Ai architecture and cost structure to include in this study [16, 17].

There is limited information showing the clinical effectiveness of AI, such as comparative effectiveness, cost-effectiveness, and other formal health technology assessments (HTA) in oncology [4, 14]. Before AI is used in clinical practice, HTA is crucial for assisting stakeholders and decision-makers in developing effective health policies and guidelines. Without clear guidelines, it is difficult for researchers, policymakers, and developers to decide if an AI system is appropriate for oncology use. Furthermore, there are still significant gaps in the literature regarding the comprehensive economic evaluation of AI in oncology. This scoping review aimed to systematically assess recent literature on health economic evaluations (HEEs) of artificial AI applications in oncology and explore the possible scope of the economic risks and benefits of integrating AI systems into oncology. By identifying key trends, disparities, and future research areas, this review will offer valuable insights for healthcare policymakers, practitioners, and researchers to enhance the economic integration of AI in oncology and promote its sustainable adoption.

Methodology

We conducted a scoping review on the economics of artificial intelligence in oncology following the Preferred Reporting Items for Systematic reviews and Meta-Analyses extension for Scoping Reviews (PRISMA-ScR) guidelines [18].

Literature search strategies

We conducted a scoping review of articles published between 1 st January 2019 to 30th June 2024, following the Arksey and O'Malley methodological framework [19]. Our sources included PubMed, Scopus, and Google Scholar. We used a combination of English keywords such as " Artificial Intelligence" or "Machine Learning " or “Deep Learning” alongside terms like " Oncology” or " Cancer" or " Neoplasms " and "Economics" or "Economic Evaluation" or "Cost–Benefit Analysis" or "Cost-Effectiveness" or "Budget Impact". Additionally, we employed the snowball strategy to identify further sources from the references of relevant full texts. Both MeSH (Medical Subject Headings) terms and free-text terms were utilized in the search.

Eligibility criteria

Studies included in this review were published between January 1, 2019, and June 30, 2024, in English. Eligible studies focused on the economic aspects of AI applications in oncology, including cost-effectiveness analyses, budget impact studies, and evaluations of economic benefits in clinical practice. This period was chosen to capture the latest advancements and impacts of AI in oncology, ensuring that the review reflects the most current developments in the field.

Data extraction

Two researchers initially screened titles and abstracts to determine if they met the inclusion criteria. After removing duplicates, the full texts of the remaining studies were reviewed to identify any potential exclusion criteria. Disagreements regarding eligibility were resolved through discussion with another team member. Data were extracted based on several variables, including the author, country of data used, study design, study population and care pathway phase (screening, diagnosis, treatment, etc.) discussed from the studies. We further extracted information related to the health economic evaluation which involved health economic analysis including economic risk and benefit, time horizon, comparator, health or cost outcomes, modelling techniques and AI model validation. Further quality assessments on the methodology of included studies were analyzed using the Philips checklist [20].

Data synthesis, analysis and quality assessment

Data were organized into an evidence matrix using a standardized template in Google Sheets. AI assisted tools were used to support the systematic review process, including Zotero 6 for reference management and duplicate removal, Rayyan for title, abstract and full text screening, and PopAI for document organization and summarization during screening. These tools applied to analyzed the 43 eligible studies following the application of predefined exclusion criteria. A critical appraisal of the studies for quality of studies was conducted using the Consolidated Health Economic Evaluation Reporting Standards for Interventions That Use Artificial Intelligence (CHEERS-AI) and the Philips checklist [20–22].

Although CHEERS was initially not developed as a quality assessment tool for health economic studies, but as a reporting checklist, its widespread use for evaluation of the design of health economic studies makes it an appropriate checklist for this scoping review. CHEERS—AI includes 38 reporting items in total, with the original 28 CHEERS-2022 checklist items with 8 elaborations to draw out potential AI-related nuances, plus 10 new AI-specific items that extend upon CHEERS-2022.Three pillars make up the Philips checklist, which was created especially to evaluate the quality of model-based HEEs: modelling approach, model data, and uncertainty assessment. The choice of a particular model type, the justification of the model parameter values, and the management of uncertainty are critical factors that influence how well the studies are regarded. Consequently, for model-based research, the Philips checklist was used in addition to the CHEERS checklist. Points were given for every satisfactory requirement. If any part of the requirement was not satisfied, a point was deducted. For every included study, a checklist score was calculated based on the percentage of the 38 satisfied criteria. The same two reviewers independently assessed the final studies included in the scoping review.

Results

Selection of source of evidence

The article selection process involved two selection phases: (1) title and abstract review, and (2) full‐text review. Figure 1 depicts the flow diagram of the study selection process. Initially, 870 records were identified through searches in three databases: PubMed (225 records), Scopus (300 records), and Google Scholar (345 records). These records underwent a screening process by an independent reviewer to remove duplicates and articles that were not relevant based on their titles and abstracts, resulting in 43 records being retained for further review [23–96]. Subsequent screening excluded 21 articles removed based on the study design. This narrowed the pool down to 22 articles that were accessed for eligibility. Finally, with the final discussion by the reviewer, an in-depth text analysis was conducted, leading to the exclusion of an additional 10 articles that did not focus on the economics of artificial intelligence in oncology, culminating in 12 studies being included in the final analysis.

Fig. 1.

Flow chart for study selection process

General overview of the included studies

The characteristics of the studies included, including their focus population and approach to AI integration, are provided in Table 1. Most of the studies reviewed were published in 2022 or later. Of the studies, there were 5 studies evaluating the economic consequences of AI use in colorectal cancer; 2 on both breast and lung cancer; 1 each on cervical and prostate cancer with another paper focused on economic review across diverse cancers such as lung and renal. Furthermore, 10 studies were concentrated on the early stage of the care pathway, such as screening, and 2 focused on the diagnostic phase. Geographically, European and America region published 5 papers respectively while only 2 Asian regions only published two papers.

Table 1.

Characteristics of the included studies (N = 12)

Authors [ref] (Year) (Location)	Title of studies	Focus cancer	Care pathway phase	Population	Study design	CHEERS-AI score, %
Mori et al., [23] (2020) (Japan)	Cost savings in colonoscopy with artificial intelligence-aided polyp diagnosis: an add-on analysis of a clinical trial (with video)	Colorectal	Diagnostic	207 patients with diminutive rectosigmoid polyps. The media age was 67 years (IQR, 59–73 years)	An add-on analysis of a clinical trial	66
Adams et al., [24] (2021) (Canada)	Development and cost analysis of a lung nodule management strategy Combining Artificial Intelligence and Lung-RADS for baseline lung cancer screening	Lung	Screening	Individuals undergoing baseline lung cancer screening, with no specific age group mentioned	Economic simulation modelling study involves retrospective secondary analysis	76
Ariea et al., [25] (2021) USA	Cost-effectiveness of artificial intelligence for screening colonoscopy: a modelling study	Colorectal	Screening	Hypothetical cohort of 100,000 individuals in the USA aged 50–100 years	Cost-effectiveness analysis using Markov Microsimulation Model	95
Mital et al., [26] (2022) (USA)	Cost-effectiveness of using artificial intelligence versus polygenic risk score to guide breast cancer screening	Breast	Screening	Women aged 40 to 74 years	Cost-effectiveness analysis using a decision-analytic model	87
Ziegelmayer et al., [27] (2022) (USA)	Cost-effectiveness of Artificial Intelligence support in computed tomography-based lung cancer screening	Lung	Screening	Individuals undergoing lung cancer screening. The specific age group is not mentioned	Markov model of Cost-effectiveness analysis	79
Shen et al., [28] (2023) (China)	Cost-effectiveness of artificial intelligence-assisted liquid-based cytology testing for cervical cancer screening in China	Cervical	Screening	Unvaccinated and unscreened women aged 30–64 years	Cost-effectiveness analysis using Markov Model, CHEERS	92
Thiruvengadam et al., [29] (2023) (USA)	An evaluation of critical factors for the cost-effectiveness of real-time computer-aided detection: sensitivity and threshold analyses using a microsimulation model	Colorectal	Screening	Study includes adults aged 45–49	Sensitivity and threshold analysis using semi-Markov microsimulation model	63
Vargas et al., [30] (2023) (UK)	Cost-effectiveness requirements for implementing artificial intelligence technology in the Women’s UK Breast Cancer Screening service	Breast	Screening	Cohort of 100,000 women from the UK, starting from the age of 19	Discrete event simulation (DES) model	82
Du et al., [31] (2024) (Sweden)	Effectiveness and cost-effectiveness of Artificial Intelligence-assisted pathology for prostate cancer diagnosis in Sweden: a microsimulation study	Prostate	Screening	Men from the age of 50 years over a lifetime horizon, specifically aged 50–74 years for screening	Microsimulation study	87
Kenseth et al., [32] (2024) (Norway)	Is risk-stratifying patients with colorectal cancer using a deep learning-based prognostic biomarker cost-effective?	Colorectal	Screening	For each group, hypothetical cohorts of 1000 patients with stage III (low risk and high risk)	Two partitioned survival models from two treatment cohort	87
Libanio et al., [33] (2024) (Netherlands, Italy, Portugal)	Combined gastric and colorectal cancer endoscopic screening may be cost-effective in Europe with the implementation of artificial intelligence: an economic evaluation	Colorectal	Screening	Men and women aged 50 to 75 years in the Netherlands, Italy and Portugal	Cost-effectiveness analysis using a Markov model	82
Reason et al., [34] (2024) (UK)	Artificial Intelligence to automate health economic modelling: a case study to evaluate the potential application of large language models	General	Diagnostic	The population of the study is not individuals but rather health economic models focused on non-small cell lung cancer (NSCLC) and renal cell carcinoma (RCC) treatments	A case study that assesses the use of artificial intelligence (AI) in automating health economic model construction	74

Health economic analysis

The most prevalent health economic analysis of the included studies was focused on cost-effectiveness (9 studies, 75%) using Markov models (5 studies, 42%). Most studies adopted the healthcare system perspective in analysis (6 studies, 50%), followed by the payer perspective (3 studies, 25%) and one study with a societal perspective. The cost-effectiveness of AI in oncology was examined in the context of medical imaging (5 studies, 42%), biomarker applications (4 studies, 33%) and health system management (3 studies, 25%). The rest of the included studies focus on cost minimization economic analysis (3 studies, 33%), with intervention, particularly in AI management lung screening strategies, large language model-based health economic evaluations, and AI-assisted diagnosis strategy for colorectal endoscopy. The time horizons adopted by the studies ranged from six months to a lifetime; one study described a horizon of less than a year, four studies described a horizon of one year, and the remaining studies utilized a horizons of three to forty years or a lifetime. Of these studies, a 3% discount rate was applied in seven studies (59%) and a 4% rate was applied in one study (8%). For a detailed overview of health economic methodologies can be found in Table 2.

Table 2.

Health economic methodology including Philip's main items of included studies (N = 12)

Authors [ref] (Year) (Location)	Economic Model	HEE type	Intervention	Compartor	Perspective	Discount rate	Time horizon	Outcome
Mori et al., [23] (2020) (Japan)	An add-on analysis of a clinical trial	CMA	AI-supported colorectal endoscopy plus diagnose-and-leave	No use of AI, resectall-polyps strategy	Payer	-	1 year	AI Cost Reductions-Japan:18.9% in colonoscopy costs, US$149.2 million; England:6.9% in costs, US$12.3 million; Norway: 7.6% in costs, US$85.2 million
Adams et al., [24] (2021) (Canada)	Economic simulation modelling	CMA	AI informed management strategy for Lung-RADS	Lung RADS only	Payer	-	6 month	$72/patient screened (Minimum net cost savings using the AI-informed management strategy)
Ariea et al., [25] (2022) (USA)	Markov model microsimulation	CEA	AI assisted colonoscopy for poly detection	Colonoscopy with and without the use of AI for polyp detection against the natural history (no screening) model	Societal	3%	5 year	CRC incidence 4.8% incremental gain discount costs 57$ per individual
Mital et al., [26] (2022) (USA)	Hybrid decision tree/microsimulation model	CEA	AI based mammography	Polygenic risk score (PRS)	Healthcare system	3%	1 year	ICER $23,755/QALLY
Ziegelmayer et al., [27] (2022) (USA)	Markov simulation model	CEA	AI assisted computed tomography based lung cancer screening	Traditional lung cancer screening	Healthcare	3%	20 year	AI support in initial screening is cost-effective up to USD1,240/patient, with a willingness-to-pay of USD100,000/QALY
Shen et al., [28] (2023) (China)	Markov simulation model, CHEERS-2022 checklist reported	CEA	AI assisted liquid based cytology	Manual LBC and HPV DNA testing	Healthcare provider’s	3%	5 year	$8790/QALY (healthcare provider)
Thiruvengadam et al., [29] (2023) (USA)	Semi-Markov microsimulation model	CEA	CAD based imaging for CRC	Traditional method	Healthcare system	3%	1 year	$579/colonoscopy (cost-effectiveness using CAD)
Vargas et al., [30] (2023) (UK)	Discrete event simulation (DES) model	CEA	AI assisted mammogram	Human radiologist	NHS system	-	3 year	Using Mia® AI as a second reader could be cost-effective for the NHS, with minimal cost and QALY difference compares to standard practice
Du et al., [31] (2024) (Sweden)	Microsimulation study	CEA	AI assisted pathology	Human pathologist	Health care	3%	Lifetime	€10 per case, < 0.001% lower QALYs
Kenseth et al., [32] (2024) Norway	Two partitioned survival models-decision tree and a partitioned survival model	CEA	Deep learning-based prognostic biomarker for CRC risk-stratification	Compare with existing clinical guidance	Health care system	4%	40 years	270,934 NOK (net monetary benefit) and 0.71 (health benefit) for stage III
Libanio et al., [33] (2024) (Netherlands, Ita;y, Portugal)	Markov model	CEA	AI assisted EDG	Stand alone EDG	Health care system	3%	1 year	With AI improving EGD accuracy by 1%, combined screening twice per decade is cost-effective in Italy, and with 96% accuracy, screening once per decade is cost-effective in the Netherlands
Reason et al., [34] (2024) (UK)	Partitioned survival models	CMA	LLM based automation of health economic evaluation	Compare drugs for cancer treatment	Payer	-	20 years for NSCLC, 40 year for RCC	For NSCLC, AI ICERs were published in USD $117,600/QALY vs. USD $117,739/QALY published for RCC, AI ICERs were CHF 107,284/QALY vs. sunitinib and CHF 105,965/QALY vs. pazopanib, compared to CHF 108,326/QALY and CHF 106,996/QALY published

HEE Health Economic Evaluation, CRC Colorectal Cancer, LBC Liquid Based Cytology, HPV Human Papilloma Virus, DNA Deoxyribonucleic Acid, CEA Cost-Effectiveness Analysis, CMA Cost-Minimization Analysis, CAD Computer Aided Diagnosis, QALY Quality Adjusted Life-Year, ICER Incremental Cost Effectiveness Ratio, LLM Large Language Model, NSCLC Non-Small Cell Lung Cancer, RCC Renal Cell Carcinoma, Lung-RADS Lung CT Screening Reporting and Data System, EGD Esophagogastroduodenoscopy, USD $ Unite States Dollar, NOK Norwegian Krone, € Euro

Quality assessment of included studies

The CHEERS—AI and Philips checklists were used to assess the methodological quality of the studies. A score was calculated as the percentage of criteria fulfilled for the analysis of the CHEERS-AI results. The scores ranged from 63 to 95%, with an average of 80%, as indicated in Table 1. The study that received the highest score used the Markov microsimulation model to examine the cost-effectiveness of artificial intelligence in colonoscopy screening [25]. Out of 12, 3 studies conducted an economic evaluation using CHEERS: Ariea et al. used the original CHEERS checklist, while Libanio et al. and Shen et al. used the updated CHEERS-2022 checklist [24, 28, 33]. Furthermore, Fig. 2, provides an overview of the studies that satisfied the CHEERS- AI checklist point ranging from 1 to 12. Although most studies provided explicit statements on the checklist, none involved an approach to patient participation and engagement. All studies mention the study population depends on the different study designs in specific locations. Only 3 studies (25%) address the distribution of impacts among distinct individuals or the implementation of modifications to align with priority populations [25, 26, 28]. Regarding the AI-related items specifically, 4 studies (33%) state that the model learning's performance will be impacted over time by the assumptions made [25, 27, 28, 34]. Six studies (50%), discuss how results can vary depending on the subgroups [25–28, 31, 34]. Six studies (50%) discuss how the AI intervention's features might raise doubts about how cost-effective it is [23, 25, 26, 30–32]. Additional information regarding the CHEERs-AI checklist can be found in Appendix 2 of the supplemental tables.

Fig. 2.

Overview of the proportion of studies reporting CHEERS-AI focus checklist items

All 12 review articles addressed economic modelling, with their characteristics summarized in Table 2. The Philips checklist was employed to evaluate the quality of these modeling-based studies, detailed in Appendix 3 of the supplemental tables. Among the reviewed studies, five types of economic modelling approaches were identified: the Markov model (N = 5), the Partitioned survival model with a decision tree (N = 3), the non-specific economic simulation (N = 2), the discrete event simulation (DES) model (N = 1), and add-on analysis of a clinical trial (N = 1). All the included studies compared the conventional approaches with AI integration and a clear model feasibility evaluation was conducted for two studies that utilized the Markov model and one used the Partitioned survival model with a decision tree [25, 28, 33]. Regarding model data, 2 studies have mentioned utility weights of the study's data and justified methods of deviation [32, 33], while only one study justified model differences using an independent dataset [32, 33]. Nevertheless, sensitivity analysis was employed in all 12 studies to report on parameter uncertainty.

Figure 2 on the CHEERS-AI checklist, several methodological domains were consistently underreported across the included studies. Patient and stakeholder engagement was absent, with both the “approach to engagement” and the “effect of engagement” items scoring zero. Among AI-specific components, modelling of AI learning over time (n = 4) and characterization of distributional effects (n = 3) were poorly addressed. Additionally, only half of the included studies focus on currency conversion methods (n = 6), characterization of heterogeneity (n = 6) and impact of AI-related uncertainty (n = 6). While reviewing Philips et al.’s quality assessment checklist for included studies (N = 12), we have reported the not applicable (NA) when the item was irrelevant to the study design, such as item "29. Has a half-cycle correction been applied to both costs and outcomes?" where most included studies were in the trial stage.

Economic benefits and risks discuss in the included studies

The included studies describe the economic benefits and risks of integrating AI in oncology, as shown in Fig. 3. From an economic benefit perspective, AI improves sensitivity and reduces healthcare costs compared to manual screening [25–28]. It streamlines workflows, lowers workload [31], minimizes recall appointments [24, 27], and optimizes treatment selection [23, 33], leading to reduced toxicity, better resource use [34], and enhanced patient outcomes [24, 26]. Furthermore, AI can significantly reduce drug waste and overall costs for patients, institutions, and insurance systems [33, 34]. However, the included studies mentioned notable economic risks, including uncertainties about AI service reimbursement under traditional fee-for-service models [25, 26, 28–30] and the need for ongoing validation to ensure long-term cost-effectiveness [27, 32]. Additional concerns involve data availability, security, fairness [30, 33], and potential costs related to errors in AI models [24, 31, 34].

Fig. 3.

Economic risk and benefit analysis of Artificial Intelligence implementation in onocolgy discussed from the included studies (N = 12)

Discussion

The scoping review systematically evaluated 12 studies, following a rigorous selection process from 870 identified records, on health economic evaluations (HEEs) of artificial intelligence (AI) integration in oncology with a focus on the economic risks and benefits. The studies varied in their focus, including colorectal, breast, lung, cervical, and prostate cancers, and evaluated cost-effectiveness using Markov models and healthcare system perspectives. Most studies assessed AI’s economic impact on early-stage care pathways, particularly screening and diagnostics, with a substantial number of studies based in High-income countries from Europe(n = 5), America(n = 5) and Asia (n = 2). Although the average CHERRs-AI score for the overall paper reaches 80%, none involved an approach to patient participation and engagement and only two studies mentioned utility weights of the study's data and justified methods of deviation. Thus, the scoping review emphasized the need for reporting using the CHEERS-AI checklist, which provides decision-makers with a comprehensive economic analysis of AI in oncology, thus improving transparency and reproducibility in the adoption of AI interventions in clinical practice [14, 20].

The included studies in the scoping review reported that AI integration in oncology is either economically cost-effective (N = 9) or cost-minimizing (N = 3) from Markov and decision trees model are used in medical imaging (N = 5), biomarker applications (N = 4), and health system management (N = 3). All the included studies compared conventional approaches with AI integration in oncology showing positive economic impact, aligning with previous research that highlights AI’s potential in radiogenomics for oncology. This potential relies on trustworthy and transparent predictive tools and algorithms to enhance disease diagnosis, prognosis, and outcome forecasting, ultimately reducing overall costs [36, 37]. Additionally, AI technology not only reduces healthcare errors, such as mislabeling drugs but also helps organizations maximize long-term returns on investment by streamlining workflows and enhancing system efficiency [10, 38, 39]. However, integrating AI into oncology presents challenges between clinical adoption and market approval, especially the economic evaluations that may be required for health insurance reimbursement. Previous reviews and current findings emphasize the need for specific regulatory frameworks to ensure comprehensive health economic evaluations while establishing cost analysis of AI in oncology to further improve beyond health outcomes [14, 40].

Additionally, the review also exposes a gap in studies from low and middle-income countries (LMIC), suggesting a need for broader geographic representation in future research. In LMICs, AI technologies in oncology can improve two main issues: addressing healthcare personnel shortages and overcoming barriers resulting from limited medical equipment [38, 43, 44]. However, the existing "digital divide" poses significant challenges, as unequal access to digital infrastructure such as appropriate hardware to run AI programs halts health equity for AI adoption in healthcare [45]. In LMICs, these AI technologies address fundamental issues on health equity: health workforce constraints, training costs, limited resource constrained systems. However, the "digital divide" and opportunity cost poses a significant barrier, as unequal access to digital infrastructure hampers effective AI implementation [97, 98]. Despite efforts to tailor technologies for LMICs, concerns about implicit bias in AI algorithms and their potential to exacerbate health inequities remain underexplored [45, 97, 98].

Additionally, AI's reliance on quantitative data and third-party classifications may lead to misrepresentation and limit the generalizability of findings across different countries [46–48]. Furthermore, the scoping review reveals that while some research addresses the distribution of AI’s impacts among distinct individuals, there remains a significant gap in patient participation and engagement in economic evaluations. According to Kumar et al., as a part of recommendations for the development of AI solutions for clinical oncology, user engagement should be addressed early on in the development and validation process to avoid downstream challenges with reimbursement [38]. This finding highlights a critical area for improvement, as more investments are needed for basic technology infrastructure with more diverse patient data from LMIC and engaging population are essential for assessing the true equity implications of AI technologies in the field of oncology. Moreover, our study findings on the absence of patient participation in the included economic evaluations raises concerns regarding health equity and real-world adoption, as it may result in limited incorporation of patient-reported outcomes, societal perspectives, and meaningful patient involvement in model development and evaluation [99, 100].

The scope of the review indicates that AI integration in oncology presents numerous economic benefits, such as increased sensitivity, lower costs, streamlined workflows, reduced workload, fewer recall appointments, optimized treatment, and enhanced patient outcomes all contributing to cost savings. In contrast, it also features economic risks, such as uncertainties around AI reimbursement, the need for long-term validation, data security concerns, and potential costs of errors in AI models. These findings echo the ethical and regulatory concerns of AI in the field of oncology which were mentioned in previous studies [45, 46, 48–50]. Although the CHEERS-AI and Philips checklists offer thorough economic evaluations of AI adoption in oncology, the scoping review highlights the need to include additional items in the checklist such as health equity, regulatory, and ethical considerations for a more comprehensive economic assessment of AI adoption in clinical oncology. Additionally, existing AI economic evaluations tend to emphasize predictive performance while underestimating or omitting real-world implementation costs such as infrastructure investment, model retraining and maintenance, regulatory compliance, and workforce training, which should be systematically incorporated in future reviews.

The scoping review has some limitations. While artificial intelligence (AI) has been developed over the past two decades, research focusing on the economic aspects of AI on oncology has been limited, which justifies our focus on articles published between January 1, 2019, and June 30, 2024. That pertinent non-English studies were overlooked is a possibility, and the exclusion of preprints and grey literature could introduce publication bias in this rapidly evolving AI field. Of 870 records identified under PRISMA-ScR guideline, 858 studies were excluded due to the absence of formal economic evaluations despite reporting AI performance metrics. This systematic approach aims to reduce selection bias in this study [18]. We discovered that a high exclusion rate resulted from the fact that many papers that were first classified as covering cost-effective AI applications lacked thorough economic analysis. Although the CHEERS-AI and Philips checklists were used to ensure that crucial elements were included in economic assessments, they did not ensure these elements were implemented correctly. While the computed score may be criticized for treating each checklist item as equally important, it assesses how thorough the economic evaluation is for each study [14]. Furthermore, the generalizability of findings may be impacted by the scarcity of studies on cost–benefit outcomes across various demographic groups.

Conclusion

The scoping review findings highlighted the exiting needs for comprehensive health economic evaluation reporting for AI integration in oncology. Despite the included studies review findings affirm that AI technologies, particularly in medical imaging, biomarker applications, and health system management, are generally cost-effective or cost-minimizing, the need for user perspectives in the development and validation of AI technologies as well as lack of perspective on health equity, regulation and ethics are needed in the adoption of technologies into actual practical. To bridge these gaps, future studies should include a more robust health-economic evaluation with a more geographically diverse population especially from low- and middle-income countries. By addressing these challenges, the transformative potential of AI can be achieved in oncology, adopting it in the healthcare system and reducing both mortality and morbidity.

Authors’ contributions

HMT was responsible for conceptualization, data curation, formal analysis, investigation, methodology, resources, software, validation, visualization, and writing the original draft. LN contributed through supervision and data validation. OAM provided supervision, data validation, and contributed to writing, reviewing, and editing the manuscript. HAR was involved in methodology development, supervision, data validation, and writing, reviewing, and editing the manuscript.

Funding

No funding was received for this study.

Data availability

No datasets were generated or analysed during the current study.

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Competing interest

The authors declare no competing interests.